Dynamic graph modification | Coursera Community
Coursera Header
Question

Dynamic graph modification

  • 1 August 2020
  • 3 replies
  • 52 views

Badge +1

Hi All,

I’m looking to build a model that uses some measures to optimise a network graph shape dynamically as the graph is trained (so the number and distribution of nodes as well as their weights). 

I understand Tensorflow 2 has made things easier but I’m looking for a guide to show me how I might do this. 

Have any of you come across anything? 

Regards,

Simon


3 replies

Badge

Hope you will find best suggestion. 

Userlevel 2
Badge +2

I’m looking to build a model that uses some measures to optimise a network graph shape dynamically as the graph is trained.

Have any of you come across anything? 

Hello Simon,

I haven’t worked with network graphs before but I’ve worked on real-time graphs from Sensor nodes [IoT]. In that I used the real-time data capturing and used it for training and predicting the future events and calculating error-function at each step. I didn’t used most of the tensorflow stuff. Just keras and statistics that I learnt.

I hope this analogy might help you with it as the major issue with your problem is grabbing that real-time network-graph data. Once you cover that milestone, after that it’s just basic stats and predictions.

Lemme know your thoughts on this.

Badge +1

Thanks Aziroazide - I don’t think I have been clear. I mean the structure of the neural network - I am looking to change it as it is trained. You may recall that in Andrew Ng’s course, we look at a fixed neural network and train it - so the values of the nodes each layer of the network change but the structure of each layer remains unchanged. Since Tensorflow compiles the shape of the network prior to running the training calculations, it was hard to make changes after that compilation. I understand that TF2 has some features that make this more plausible though. 

Please do let me know if I’m still not being clear.

Simon

Reply